The paper by Jeffrey L. Elman explores the interaction between learning and development in neural networks, particularly in the context of language acquisition. Elman argues that the most significant learning occurs during childhood, a period marked by dramatic maturational changes. He investigates this relationship using connectionist networks, which are trained to process complex sentences involving relative clauses, number agreement, and verb argument structures. The results show that training fails when networks are fully formed and 'adult-like' in their capacity. However, training succeeds only when networks start with limited working memory and gradually 'mature' to the adult state. This suggests that developmental restrictions on resources may be necessary for mastering complex domains, and that successful learning depends on starting small. Elman's findings contrast with other studies in the connectionist literature, where it is known that some problems can be learned best when the entire data set is available to the network. However, this paper argues that in the context of language learning, starting small can be beneficial. The author explains that the initial limitations on memory capacity act as a filter, focusing learning on a subset of facts that form the foundation for future success. This early commitment to basic grammatical factors constrains the solution space, reducing the risk of false solutions and improving learning performance. The paper also discusses the statistical basis for learning, the representation of experience, constraints on new hypotheses, and how the ability to learn changes over time. These properties collectively limit the power of networks, but when embedded in a system that develops over time (starts small), they compensate for these limitations. Elman's findings have implications for understanding the interaction between learning and development in humans, suggesting that developmental changes may enable more effective learning.The paper by Jeffrey L. Elman explores the interaction between learning and development in neural networks, particularly in the context of language acquisition. Elman argues that the most significant learning occurs during childhood, a period marked by dramatic maturational changes. He investigates this relationship using connectionist networks, which are trained to process complex sentences involving relative clauses, number agreement, and verb argument structures. The results show that training fails when networks are fully formed and 'adult-like' in their capacity. However, training succeeds only when networks start with limited working memory and gradually 'mature' to the adult state. This suggests that developmental restrictions on resources may be necessary for mastering complex domains, and that successful learning depends on starting small. Elman's findings contrast with other studies in the connectionist literature, where it is known that some problems can be learned best when the entire data set is available to the network. However, this paper argues that in the context of language learning, starting small can be beneficial. The author explains that the initial limitations on memory capacity act as a filter, focusing learning on a subset of facts that form the foundation for future success. This early commitment to basic grammatical factors constrains the solution space, reducing the risk of false solutions and improving learning performance. The paper also discusses the statistical basis for learning, the representation of experience, constraints on new hypotheses, and how the ability to learn changes over time. These properties collectively limit the power of networks, but when embedded in a system that develops over time (starts small), they compensate for these limitations. Elman's findings have implications for understanding the interaction between learning and development in humans, suggesting that developmental changes may enable more effective learning.